Knee MR imaging is indispensable for the diagnosis of ligament tears and meniscal injuries. However, manual knee MR slice positioning is time-consuming and has poor reproducibility for follow-up studies.

Previous studies reported automatic slice positioning algorithms based on anatomical landmark detection in rough 3D scout images^1,2. The Active shape model based method¹ detects landmarks by accounting for the appearance of the femur and tibia; however, it is not robust to severe artifacts. Zhan’s method² improves landmark detection robustness by using a redundant and hierarchical learning method; however, these landmarks do not necessarily yield desired slice positions due to inter-patient variability. For an MR slice positioning centered on the femoral condyle in a sagittal plan in the left-right direction (Fig. 1A-a), the desired slice positioning (yellow FOV) may not agree with anatomical landmarks, as landmarks are not necessarily equivalent to FOV tangent points. For an MR slice positioning based on the angles of the femur and tibia (Fig. 1B-s), at least two landmark points must be detected per bone; however, 3D scout image coverage varies from scan to scan, especially in the head-foot direction, so that landmarks may be cut-off. For these reasons, methods which are entirely dependent on anatomical landmark detection are not robust.

Hence, we propose an automatic slice positioning algorithm using not only machine-learning based anatomical landmark detection, but also image processing based femoral contour detection and bone angle estimation. Our proposed method can accurately, quickly and reproducibly position knee MR.

Data acquisition

3D fast field echo (FFE) scout images covering both the left and right knee were acquired from 50 healthy volunteers using 3T MRI scanner with FOV = 500mm x 500mm x 160mm in less than 25 seconds.

Data processing

The proposed method consists of three steps: target knee detection, anatomical landmark detection and image processing. First, the target knee is extracted using discriminant analysis from the scout image which includes both knees.

Then, anatomical landmarks are detected using extremely randomized trees method³ which is fast with high performance of classification. Positional relationship between landmarks is used to improve robustness against image artifacts. Representative landmarks are shown in Fig. 2. Detected landmarks are used in subsequent image processing.

Finally, image processing is applied to yield final slice positioning. For slice positioning in the sagittal plan, the femoral condyle contour is detected by solving the shortest path problem between landmarks. The cost of path is defined by image intensity and gradients. The slice positioning center in the left-right direction is calculated as the center of the rectangle bounding the computed contours. For slice positioning based on femur and tibia angles, an energy maximization function is used to independently estimate angles for each bone. For femur angle estimation, the energy $$$E(\theta)$$$ inside a local region $$${\bf R}(\theta)$$$ just above the detected femoral center point is calculated. The local region $$${\bf R}(\theta)$$$ can be at any angle $$$\theta$$$ from the femoral center point. The energy function is $$E(\theta)=\sum_{{\bf R}(\theta)}I_{mag}\cos(2(I_{dir}-(\theta+90))),$$ where $$$I_{mag}$$$ and $$$I_{dir}$$$ denotes the magnitude and direction of the intensity gradient inside the local region $$${\bf R}(\theta)$$$, respectively. Energy is maximized so that the femur direction and intensity gradient are normal. The tibia angle is estimated in a similar manner using the detected tibial center point.

Our experiments uses 50 scout images (34 for training, 16 for testing). Ground truths were based on three technologists. To evaluate the proposed automatic slice positioning algorithm, we compared inter- technologist error with computed errors. Mean and standard deviation of translational and rotational errors are shown in Fig. 3. The processing time was approximately 1.0 seconds on a 3.5 GHz CPU. Our method achieved an accuracy comparable to inter- technologist errors while maintaining faster processing time than previous methods ([1]: 15sec and [2]: 5sec for different datasets, slice positioning and CPUs).

Additionally, to evaluate intra-patient reproducibility, 6 volunteers were scanned multiple times and slice positioning results were compared. Figure 4 shows that our method can yield consistent intra-patient slice positioning regardless of scan-to-scan variations (Fig. 4, Scan A vs. B).

We proposed an automatic knee MR slice positioning algorithm using machine-learning based landmark detection combined with image processing based methods, such as femoral contour detection and angle estimation of the femur and tibia. Experimental results showed that our method was accurate, fast and reproducible, with potential benefits for technologists, doctors and patients by improving examination workflow.